Technology of improving the land use data accuracy for geospatial modeling tasks

• Date: 2026
• Main Authors: Okhariev, Viacheslav, Pidsadnii, Serhii
• Format: Article
• Language: Ukrainian
• Published: Kyiv National University of Construction and Architecture 2026
• Subjects: geospatial modeling; land use data; machine learning
• Online Access: https://es-journal.in.ua/article/view/364955
• Journal Title: Environmental safety and natural resources

The aim of the research is increasing the accuracy of land cover classification (LULC) based on Dynamic World (DW) data for the controlled territory of Ukraine. One of the key problems of such data is the regional specificity of the land use regime, which reduces the reliability of the global average world classification. The basic accuracy of the DW classification for all of Ukraine is 51%, which is significantly lower than the declared one (~72%). The exclusion of occupied and frontline territories and partial re-verification of reference points allowed to increase the accuracy to 64.58%, Kappa coefficient (κ) 0.600. Methods for refining the classification based on the seasonal amplitude of NDVI and the summer phenology filter did not provide an increase in accuracy due to the significant spectral overlap of vegetation cover classes. The method of object-oriented analysis based on cadastral boundaries showed the inaccuracy of such an approach even at the stage of material preparation. Instead, the Random Forest machine learning model, built on NDVI phenological indicators, annual mean class, and DW probability bands, achieved an accuracy of 68.10%, Kappa 0.637, which corresponds to the category of substantial agreement and is the best result among the studied methods.